Summary

• Bull will supply HLRE-5 and five years of support under a €40 million contract.
• The system is designed to run two global climate simulations simultaneously at five-kilometre resolution.
• Direct liquid cooling and increased heat recovery support a target of 75% lower energy consumption per simulated year.

Germany’s Climate Computing Centre (DKRZ) has awarded Bull a €40 million contract for its next supercomputer, combining higher-resolution climate modelling with AI research and a planned substantial reduction in the energy required for scientific simulations.

The High-performance Computing System for Earth System Research, known as HLRE-5, will be based on Bull’s BullSequana XH3500 platform and is expected to become available to researchers in early 2028. The contract covers the computing system, storage infrastructure and five years of support.

DKRZ says the installation will be capable of running two global climate simulations simultaneously at five-kilometre spatial resolution, allowing researchers to investigate complex Earth system behaviour at a finer level of detail than earlier computing generations could support economically.

The system is also designed to reduce energy consumption per simulated year by approximately 75% compared with Levante, the centre’s current supercomputer. That figure is a projected improvement in the efficiency of scientific workloads, rather than a commitment to reduce the facility’s total electricity consumption by the same percentage.

Funding will come principally from the Helmholtz Association, the Max Planck Society and the City of Hamburg, with additional German research investment allowing the system’s planned capacity to be increased.

Greater detail changes computing requirements

Climate models divide the Earth into computational regions and calculate interactions between the atmosphere, oceans, land and other components of the climate system. Increasing geographical resolution enables researchers to represent smaller physical features, but also adds considerable computational work.

More grid points can require shorter simulation intervals and larger volumes of data, meaning an apparently modest increase in resolution may produce a disproportionately large increase in processing requirements.

Higher spatial resolution does not automatically produce more accurate findings. The representation of physical processes, quality of observations and treatment of uncertainty remain important, even when a computer can calculate environmental conditions over smaller areas.

HLRE-5 is intended to increase the amount of detailed modelling DKRZ can undertake while supporting additional methods that combine established physics-based simulations with machine learning.

AI techniques can contribute to selected modelling and analytical tasks, but they must be evaluated against scientific observations and conventional methods. Their usefulness will depend on whether they can produce reliable results at a reasonable computational cost.

The planned ability to run two global simulations at five-kilometre resolution simultaneously is particularly relevant where researchers need to compare different assumptions or conduct several scientific investigations within limited computing time.

Researchers will still need to adapt software to the new hardware and verify that results remain scientifically consistent after applications move to a different computing architecture.

Liquid cooling supports denser infrastructure

Bull’s direct liquid cooling technology will circulate water through computing nodes to remove heat generated by processors and associated components. The system uses a closed circuit that can transfer heat more efficiently than arrangements relying exclusively on air cooling.

Liquid cooling is increasingly relevant to high-performance computing and AI infrastructure because growing processor density concentrates substantial electrical demand within individual cabinets.

Effective cooling influences the number of computing devices that can operate within a given space, the electricity required by supporting equipment and the design of the building accommodating the installation.

DKRZ already transfers some waste heat from Levante to the neighbouring chemistry department at the University of Hamburg. The new installation will involve improvements to the cooling infrastructure, making a larger proportion of the recovered heat available for useful purposes.

The university is considering extending the arrangement to additional buildings around its KlimaCampus, potentially allowing the computing facility to contribute more energy to local heating requirements.

Heat recovery depends on several practical conditions. The temperature and availability of the recovered energy must match the requirements of suitable users, while infrastructure is needed to transfer it between buildings.

Consequently, the theoretical amount of heat available from a data centre does not necessarily equal the quantity that can be used economically. Seasonal demand, operating schedules and the cost of connecting consumers all affect the value of a recovery scheme.

DKRZ’s established arrangement with the university provides a practical starting point for expansion, although the centre has not quantified the additional heat expected to be recovered from HLRE-5.

Efficiency requires a meaningful comparison

The projected 75% reduction in energy per simulated year is an important qualification in assessing the project’s environmental performance.

A more powerful supercomputer may carry out many additional simulations and use substantial electricity while producing scientific results more efficiently than its predecessor. Total power demand and useful scientific work therefore need to be considered separately.

The installation’s actual efficiency will depend on software performance, utilisation, operating conditions and the characteristics of the simulations selected for comparison.

Direct liquid cooling may also improve infrastructure efficiency, but that improvement does not establish the environmental impact of the entire computing operation. Electricity generation, equipment manufacture and the broader useful life of the installation contribute to the overall assessment.

These distinctions are increasingly important as AI infrastructure operators and scientific computing centres compete for electricity, specialised processors and engineering resources.

Large commercial AI facilities are often evaluated using measures such as installed power capacity, while research centres must also consider the quantity and scientific quality of work their systems deliver.

HLRE-5’s design seeks to improve that relationship by providing additional computing capacity without proportionately increasing the energy required for equivalent scientific calculations.

Support and migration remain significant costs

The five-year support agreement recognises that the cost of operating a supercomputer extends well beyond the initial delivery of hardware.

Scientific software can have long development histories and depend on specialised programming methods. Moving applications to new processor and memory architectures may require extensive engineering work before researchers obtain the expected performance benefits.

Storage systems must also handle large simulation outputs reliably, while operations teams manage access, maintenance, technical failures and the competing requirements of different research programmes.

Research computing procurement consequently involves long-term technical compatibility as well as immediate processing capacity. A system that performs well in a vendor benchmark may deliver less practical benefit if important scientific applications cannot exploit its architecture efficiently.

The 2028 service date allows time for installation and preparation, but also places the project within an infrastructure market characterised by changing processor technology and competing demand for highly specialised computing components.

DKRZ will need to validate the installation against real research workloads after commissioning. Its projected modelling capacity and efficiency improvements will become measurable once climate applications are operating on the completed system.

The €40 million contract secures the next stage of Germany’s climate computing infrastructure, with the research benefits dependent on successful software migration, scientific validation and the actual efficiency achieved under sustained operation.